Can an AI agent manage Meta ads for a small eCommerce brand?
Yes, with limits. An AI agent can read your Meta account every day, report what is working, draft new creatives and propose budget changes. Spending more, launching campaigns and changing offers should wait for a person to approve. The bigger job is checking Meta's numbers against your real orders. In our account, Meta credited subscription renewals to ads, which made weak campaigns look profitable.
From July through October 2026, Mully spent $18,570 on Meta ads. Meta reported 123 purchases and a ROAS of 1.49. Our books, which count first-time buyers separately, put new-customer ROAS across all ads at about 1.25. Both numbers came from the same months. The gap between them is most of what an AI ad agent is for. The month-by-month breakdown is in Blended vs new-customer ROAS: which should an eCommerce brand track?.
This guide covers what we let our Meta agents do, what we do not, and the checks that keep their reports honest.
Can an AI agent run Meta ads on its own?
It can, but it should not start that way. We split the work across two agents with different permission levels:
| Agent | What it does | Level | KPI |
|---|---|---|---|
| Ad Creative | Drafts images, video and copy | L1, drafts only | Creatives beating our CTR benchmark each month |
| Ad Buyer | Reads results, proposes budget and campaign changes | L2, acts after approval | Blended ROAS at or above 2.0 |
L1 means it drafts and reports. L2 means it can make the change, but only after a person approves. An agent moves up a level after hitting its KPI for 4 straight weeks with no incidents, and drops a level after any unapproved action.
That last rule came from a real one. In October, a campaign called Teb1 ran without approval. It spent $1,893 and recorded zero purchases before it was paused. Since then, our plan says every agent that can touch spend needs a cap, an approval level and a kill switch.
How do you set up a clean test for an agent to read?
Turn off the parts of Meta that change your ads for you. Meta says Advantage+ creative enhancements may adjust the media and text you upload, and that some are on by default (Meta Business Help Center). We turn them all off on test ads, so the ad that runs is the ad we designed. Otherwise the agent cannot tell whether a headline won or Meta's crop and overlay did.
Then pick the right number for the stage:
- Early, few purchases: judge on CTR and CPC. A handful of purchases is noise.
- Once purchases show up: switch to ROAS at the product level.
- When an ad gets starved: if Meta gives one ad almost no spend, move it to its own budget instead of breaking the campaign that works. Our "Shop | Sales | Product Test | ABO" campaign exists for exactly that.
Meta's own guidance explains why patience matters: an ad set needs around 50 optimized conversion events to exit the learning phase (Meta Business Help Center). Most small-brand tests never reach that, so the agent should say "not enough data" more often than "winner". For a worked example of a headline-only test and why the click rate winner was not the cheapest visit, see how to test ad headlines without changing the image.
Why can't you trust Meta's ROAS on its own?
Because Meta credits any purchase that fits its attribution window, including ones the ad did not cause. Meta's default setting counts conversions within 7 days of a click, 1 day of a view and 1 day of an engagement (Meta Business Help Center).
For a subscription brand, that is a trap. When our agent read the October product test, most of the orders Meta credited were $266 to $280, which is our subscription renewal amount, during renewal week. Those members were going to renew anyway. The agent's call was to read that campaign as clicks, not revenue.
The fix is a weekly check against your own orders:
- Compare Meta's purchase count to new-customer orders in Shopify for the same days.
- Flag any credited order that matches a renewal or subscription price.
- Report two ROAS numbers: Meta's and new-customer ROAS from your books.
What did the agent find was actually working?
A clear reason to buy, more than any discount. From May to late October, our best Meta campaigns were:
| Campaign | Spend | Purchases | ROAS |
|---|---|---|---|
| Subscription offer with a free rangefinder | $4,941 | 95 | 3.78 |
| "Not a discount box" copy test | $3,284 | 30 | 2.13 |
| Shop ads, single hero products | about $1,700 to $2,600 each | 7 to 13 | about 1.0 to 1.1 |
The shop ads had the best click rates of the year, but ROAS sat around 1.0. People liked the products and did not buy at the same rate. That is the kind of split an agent should state plainly instead of rounding up.
How do you keep an ad agent from wasting money?
Put the brakes inside Meta, not just inside the agent. Meta's automated rules check campaigns, ad sets and ads and can turn them off or change budgets when conditions are met. They apply at the ad account level, so they work no matter who on your team made the ad (Meta Business Help Center).
A simple setup:
- A rule that turns off any ad set above your maximum cost per purchase after a set spend.
- A spend notification on every campaign.
- The agent proposes, a person approves, then the change is made.
- One weekly report that lists every change, who approved it and what happened after.
Do not let the agent spend more than you would let a new hire spend in their first month.
Common questions
Should the agent use Meta's automated rules or its own logic?
Both. Meta's automated rules are a good kill switch because they run inside the ad account no matter who launched the ad. The agent's job is the judgment work rules cannot do, like checking reported purchases against real orders and explaining why a campaign should change.
How long should a test run before the agent calls it?
Longer than most people wait. Meta says an ad set needs around 50 optimized conversion events to exit the learning phase. A small brand may never get there on one test, so we read early tests on CTR and CPC and only switch to ROAS once purchases show up.
Can the agent write the ad copy too?
Yes, as drafts. Our creative agent runs at L1, so it produces creatives and copy for a person to approve. The KPI is how many of its creatives beat our CTR benchmark each month, which keeps it honest about whether its drafts are actually better.
What spend limit should an AI ad agent have?
Zero until it has earned trust. Start it as a reporter, then let it make changes after approval. If you later let it act on its own, give it a hard daily cap and an automated rule that turns ads off above a cost you set. Any unapproved action should drop it back a level.